libmach: fix new thread race with Linux

If you look at the sequence of values returned
by waitpid, it simply tells us about the child of
clone before it tells us that the parent called clone.
There's nothing we can do but assume unexpected
tids are newly cloned children.

Tested with 6prof on godoc.

Fixes #251.

R=r
CC=golang-dev
https://golang.org/cl/2167045

src/libmach/linux.c

@@ -202,6 +202,8 @@ attachthread(int pid, int tid, int *new, int newstate)
         memset(t, 0, sizeof *t);
 
 	thr[nthr++] = t;
+	if(pid == 0 && nthr > 0)
+		pid = thr[0]->pid;
 	t->pid = pid;
 	t->tid = tid;
 	t->state = newstate;
@@ -296,7 +298,9 @@ wait1(int nohang)
 	if(nohang != 0)
 		nohang = WNOHANG;
 
+	status = 0;
 	tid = waitpid(-1, &status, __WALL|WUNTRACED|WSTOPPED|WCONTINUED|nohang);
+
 	if(tid < 0)
 		return -1;
 	if(tid == 0)
@@ -305,11 +309,15 @@ wait1(int nohang)
 	if(trace > 0 && status != NormalStop)
 		fprint(2, "TID %d: %#x\n", tid, status);
 
-	// If we've not heard of this tid, something is wrong.
 	t = findthread(tid);
 	if(t == nil) {
-		fprint(2, "ptrace waitpid: unexpected new tid %d status %#x\n", tid, status);
-		return -1;
+		// Sometimes the kernel tells us about new threads
+		// before we see the parent clone.
+		t = attachthread(0, tid, &new, Stopped);
+		if(t == nil) {
+			fprint(2, "failed to attach to new thread %d\n", tid);
+			return -1;
+		}
 	}
 
 	if(WIFSTOPPED(status)) {
@@ -339,8 +347,6 @@ wait1(int nohang)
 				}
 				t->child = data;
 				attachthread(t->pid, t->child, &new, Running);
-				if(!new)
-					fprint(2, "ptrace child: not new\n");
 				break;
 
 			case PTRACE_EVENT_EXEC: